VLMs for Robotics

Enabling reliable long-horizon robotic manipulation is a crucial step toward open-world embodied intelligence. However, VLM-based planners treat each step as an isolated observation-to-action mapping, forcing them to reinfer scene geometry from raw pixels at every decision point while remaining unaware of how prior actions have reshaped the environment. Despite strong short-horizon performance, these systems lack the spatio-temporal reasoning required for persistent geometric anchoring and memory of action-triggered state transitions. Without persistent state tracking, perceptual errors accumulate across the execution horizon, temporarily occluded objects are catastrophically forgotten, and these compounding failures lead to precondition violations that cascade through subsequent steps. In contrast, humans maintain a persistent mental model that continuously tracks spatial relations and action consequences across interactions rather than reconstructing them at each instant. Inspired by this human capacity for causal spatio-temporal reasoning with persistent memory, we propose RoboStream, a training-free framework that achieves geometric anchoring through Spatio-Temporal Fusion Tokens (STF-Tokens), which bind visual evidence to 3D geometric attributes for persistent object grounding, and maintains causal continuity via a Causal Spatio-Temporal Graph (CSTG) that records action-triggered state transitions across steps. This design enables the planner to trace causal chains and preserve object permanence under occlusion without additional training or fine-tuning. RoboStream achieves 90.5% on long-horizon RLBench and 44.4% on challenging real-world block-building tasks, where both SoFar and VoxPoser score 11.1%, demonstrating that spatio-temporal reasoning and causal memory are critical missing components for reliable long-horizon manipulation.

VLA architectures that pair a pretrained VLM with a flow-matching action expert have emerged as a strong paradigm for language-conditioned manipulation. Yet the VLM, optimized for semantic abstraction and typically conditioned on static visual observations, tends to attenuate fine-grained geometric cues and often lacks explicit temporal evidence for the action expert. Prior work mitigates this by injecting auxiliary visual features, but existing approaches either focus on static spatial representations or require substantial architectural modifications to accommodate temporal inputs, leaving temporal information underexplored. We propose Plug-in Visual Injection (PVI), a lightweight, encoder-agnostic module that attaches to a pretrained action expert and injects auxiliary visual representations via zero-initialized residual pathways, preserving pretrained behavior with only single-stage fine-tuning. Using PVI, we obtain consistent gains over the base policy and a range of competitive alternative injection strategies, and our controlled study shows that temporal video features (V-JEPA2) outperform strong static image features (DINOv2), with the largest gains on multi-phase tasks requiring state tracking and coordination. Real-robot experiments on long-horizon bimanual cloth folding further demonstrate the practicality of PVI beyond simulation.

Recent advances in robot manipulation increasingly leverage Vision-Language Models (VLMs) for high-level reasoning, such as decomposing task instructions into sequential action plans expressed in natural language that guide downstream low-level motor execution. However, current benchmarks do not assess whether these plans are spatially executable, particularly in specifying the exact spatial locations where the robot should interact to execute the plan, limiting evaluation of real-world manipulation capability. To bridge this gap, we define a novel task of grounded planning and introduce GroundedPlanBench, a newly curated benchmark for spatially grounded long-horizon action planning in the wild. GroundedPlanBench jointly evaluates hierarchical sub-action planning and spatial action grounding (where to act), enabling systematic assessment of whether generated sub-actions are spatially executable for robot manipulation. We further introduce Video-to-Spatially Grounded Planning (V2GP), an automated data generation framework that leverages real-world robot video demonstrations to improve spatially grounded long-horizon planning. Our evaluations reveal that spatially grounded long-horizon planning remains a major bottleneck for current VLMs. Our results demonstrate that V2GP provides a promising approach for improving both action planning and spatial grounding performance, validated on our benchmark as well as through real-world robot manipulation experiments, advancing progress toward spatially actionable planning.

Recent world-model-based Vision-Language-Action (VLA) architectures have improved robotic manipulation through predictive visual foresight. However, dense future prediction introduces visual redundancy and accumulates errors, causing long-horizon plan drift. Meanwhile, recent sparse methods typically represent visual foresight using high-level semantic subtasks or implicit latent states. These representations often lack explicit kinematic grounding, weakening the alignment between planning and low-level execution. To address this, we propose StructVLA, which reformulates a generative world model into an explicit structured planner for reliable control. Instead of dense rollouts or semantic goals, StructVLA predicts sparse, physically meaningful structured frames. Derived from intrinsic kinematic cues (e.g., gripper transitions and kinematic turning points), these frames capture spatiotemporal milestones closely aligned with task progress. We implement this approach through a two-stage training paradigm with a unified discrete token vocabulary: the world model is first trained to predict structured frames and subsequently optimized to map the structured foresight into low-level actions. This approach provides clear physical guidance and bridges visual planning and motion control. In our experiments, StructVLA achieves strong average success rates of 75.0% on SimplerEnv-WidowX and 94.8% on LIBERO. Real-world deployments further demonstrate reliable task completion and robust generalization across both basic pick-and-place and complex long-horizon tasks.

We introduce $Ψ_0$ (Psi-Zero), an open foundation model to address challenging humanoid loco-manipulation tasks. While existing approaches often attempt to address this fundamental problem by co-training on large and diverse human and humanoid data, we argue that this strategy is suboptimal due to the fundamental kinematic and motion disparities between humans and humanoid robots. Therefore, data efficiency and model performance remain unsatisfactory despite the considerable data volume. To address this challenge, \ours;decouples the learning process to maximize the utility of heterogeneous data sources. Specifically, we propose a staged training paradigm with different learning objectives: First, we autoregressively pre-train a VLM backbone on large-scale egocentric human videos to acquire generalizable visual-action representations. Then, we post-train a flow-based action expert on high-quality humanoid robot data to learn precise robot joint control. Our research further identifies a critical yet often overlooked data recipe: in contrast to approaches that scale with noisy Internet clips or heterogeneous cross-embodiment robot datasets, we demonstrate that pre-training on high-quality egocentric human manipulation data followed by post-training on domain-specific real-world humanoid trajectories yields superior performance. Extensive real-world experiments demonstrate that \ours\ achieves the best performance using only about 800 hours of human video data and 30 hours of real-world robot data, outperforming baselines pre-trained on more than 10$\times$ as much data by over 40% in overall success rate across multiple tasks. We will open-source the entire ecosystem to the community, including a data processing and training pipeline, a humanoid foundation model, and a real-time action inference engine.

Embodied task planning demands vision-language models to generate action sequences that are both visually grounded and causally coherent over time. However, existing training paradigms face a critical trade-off: joint end-to-end training often leads to premature temporal binding, while standard reinforcement learning methods suffer from optimization instability. To bridge this gap, we present Staged Vision-Language Learning (SVLL), a unified three-stage framework for robust, physically-grounded embodied planning. In the first two stages, SVLL decouples spatial grounding from temporal reasoning, establishing robust visual dependency before introducing sequential action history. In the final stage, we identify a key limitation of standard Direct Preference Optimization (DPO), its purely relative nature -- optimizing only the preference gap between winning and losing trajectories while neglecting absolute likelihood constraints on optimal path, often yields unsafe or hallucinated behaviors. To address this, we further introduce Bias-DPO, a novel alignment objective that injects an inductive bias toward expert trajectories by explicitly maximizing likelihood on ground-truth actions while penalizing overconfident hallucinations. By anchoring the policy to the expert manifold and mitigating causal misalignment, SVLL, powered by Bias-DPO, ensures strict adherence to environmental affordances and effectively suppresses physically impossible shortcuts. Finally, extensive experiments on the interactive AI2-THOR benchmark and real-world robotic deployments demonstrate that SVLL outperforms both state-of-the-art open-source (e.g., Qwen2.5-VL-7B) and closed-source models (e.g., GPT-4o, Gemini-2.0-flash) in task success rate, while significantly reducing physical constraint violations.

Vision-Language-Action (VLA) systems have shown strong potential for language-driven robotic manipulation. However, scaling them to long-horizon tasks remains challenging. Existing pipelines typically separate data collection, policy learning, and deployment, resulting in heavy reliance on manual environment resets and brittle multi-policy execution. We present RoboClaw, an agentic robotics framework that unifies data collection, policy learning, and task execution under a single VLM-driven controller. At the policy level, RoboClaw introduces Entangled Action Pairs (EAP), which couple forward manipulation behaviors with inverse recovery actions to form self-resetting loops for autonomous data collection. This mechanism enables continuous on-policy data acquisition and iterative policy refinement with minimal human intervention. During deployment, the same agent performs high-level reasoning and dynamically orchestrates learned policy primitives to accomplish long-horizon tasks. By maintaining consistent contextual semantics across collection and execution, RoboClaw reduces mismatch between the two phases and improves multi-policy robustness. Experiments in real-world manipulation tasks demonstrate improved stability and scalability compared to conventional open-loop pipelines, while significantly reducing human effort throughout the robot lifecycle, achieving a 25% improvement in success rate over baseline methods on long-horizon tasks and reducing human time investment by 53.7%.

Social robot navigation requires a sophisticated integration of scene semantics and human social norms. Scaling up Vision Language Models (VLMs) generally improves reasoning and decision-making capabilities for socially compliant navigation. However, increased model size incurs substantial computational overhead, limiting suitability for real-time robotic deployment. Conversely, lightweight VLMs enable efficient inference but often exhibit weaker reasoning and decision-making performance in socially complex environments. Achieving both strong reasoning ability and efficiency remains an open challenge. To bridge this gap, we propose Group Competitive Learning (GCL), a strategy designed to amplify the capabilities of lightweight VLMs. Our strategy introduces the Group Competitive Objective (GCO) to harmonize global semantics with distributional regularization, alongside Asymmetric Group Optimization (AGO) to explore the upper limits of model performance. Empirical evaluations on social navigation benchmarks demonstrate that GCL significantly elevates VLM performance. Specifically, GCL enables the Qwen2.5-VL-3B learner model and guide Qwen3-VL-4B to achieve an F1 score of 0.968 and 0.914, representing 40% and 12% improvement over vanilla supervised fine-tuning (SFT). Notably, under vanilla SFT, the 3B model initially trails the 8B model (F1: 0.692 vs. 0.755). However, through the GCL, the 3B model outperforms (28%) the 8B baseline model. These results suggest that GCL provides an effective solution for achieving both high accuracy and computational efficiency in real-world deployment.

Robots are increasingly expected to execute open ended natural language requests in human environments, which demands reliable long horizon execution under partial observability. This is especially challenging for humanoids because locomotion and manipulation are tightly coupled through stance, reachability, and balance. We present a humanoid agent framework that turns VLM plans into verifiable task programs and closes the loop with multi object 3D geometric supervision. A VLM planner compiles each instruction into a typed JSON sequence of subtasks with explicit predicate based preconditions and success conditions. Using SAM3 and RGB-D, we ground all task relevant entities in 3D, estimate object centroids and extents, and evaluate predicates over stable frames to obtain condition level diagnostics. The supervisor uses these diagnostics to verify subtask completion and to provide condition-level feedback for progression and replanning. We execute each subtask by coordinating humanoid locomotion and whole-body manipulation, selecting feasible motion primitives under reachability and balance constraints. Experiments on tabletop manipulation and long horizon humanoid loco manipulation tasks show improved robustness from multi object grounding, temporal stability, and recovery driven replanning.

Vision-Language-Action (VLA) models enable generalist robotic manipulation but suffer from high inference latency. This bottleneck stems from the massive number of visual tokens processed by large language backbones. Existing methods either prune or merge tokens uniformly, degrading the spatial reasoning essential for robotic control. We present DepthCache, a training-free framework that leverages depth as a structural prior for visual token compression. It partitions observations into depth-based regions and applies spatially differentiated merge ratios, preserving the near-field workspace while compressing the distant background. To exploit temporal redundancy, DepthCache distributes the merging process across consecutive frames, ensuring consistent representations while reducing per-step computation. A motion-adaptive pipeline further optimizes auxiliary view compression based on end-effector dynamics. The framework requires no model modification, generalizing across diverse VLA architectures. On the LIBERO benchmark, DepthCache achieves up to 1.28x inference speedup with less than 1% average success rate degradation across three VLA models (pi_0.5, OpenVLA, GR00T), whereas pruning and merging baselines incur 4--24% degradation at comparable compression. Real-world experiments on a physical manipulator demonstrate that DepthCache enables faster task throughput and more responsive closed-loop control in latency-sensitive scenarios.

World Models (WMs) offer a promising mechanism for post-training Vision-Language-Action (VLA) policies by providing dynamics priors that improve generalization under task and scene variation. However, most WM-based post-training methods rely on pixel-space supervision, making policies sensitive to visual artifacts introduced by imperfect WM rollouts. We present World2Act, a latent-space post-training framework that transfers WM dynamics to the VLA policy without pixel-space supervision. World2Act operates in two stages: 1) it induces a shared video-action latent space by contrastively aligning WM-dynamics latents with action embeddings, and 2) it post-trains the VLA by guiding policy action representations toward WM-imagined dynamics rather than decoded pixels. Built on GR00T-N1.6, World2Act delivers absolute success-rate gains of up to +2.5% on simulation benchmarks (RoboCasa, LIBERO, Bridge-SIMPLER) and +6.7% on a real robot over finetuned VLA baselines. Notably, it outperforms pixel-space WM supervision by up to +6.0%, including on LIBERO where pixel supervision degrades the baseline, suggesting that latent WM dynamics offer a more stable WM-based post-training alternative to pixel-space transfer.

Language-conditioned local navigation requires a robot to infer a nearby traversable target location from its current observation and an open-vocabulary, relational instruction. Existing vision-language spatial grounding methods usually rely on vision-language models (VLMs) to reason in image space, producing 2D predictions tied to visible pixels. As a result, they struggle to infer target locations in occluded regions, typically caused by furniture or moving humans. To address this issue, we propose BEACON, which predicts an ego-centric Bird's-Eye View (BEV) affordance heatmap over a bounded local region including occluded areas. Given an instruction and surround-view RGB-D observations from four directions around the robot, BEACON predicts the BEV heatmap by injecting spatial cues into a VLM and fusing the VLM's output with depth-derived BEV features. Using an occlusion-aware dataset built in the Habitat simulator, we conduct detailed experimental analysis to validate both our BEV space formulation and the design choices of each module. Our method improves the accuracy averaged across geodesic thresholds by 22.74 percentage points over the state-of-the-art image-space baseline on the validation subset with occluded target locations. Our project page is: https://xin-yu-gao.github.io/beacon.

As robotic systems execute increasingly difficult task sequences, so does the number of ways in which they can fail. Video Anomaly Detection (VAD) frameworks typically focus on singular, low-level kinematic or action failures, struggling to identify more complex temporal or spatial task violations, because they do not necessarily manifest as low-level execution errors. To address this problem, the main contribution of this paper is a new VAD-inspired architecture, TIMID, which is able to detect robot time-dependent mistakes when executing high-level tasks. Our architecture receives as inputs a video and prompts of the task and the potential mistake, and returns a frame-level prediction in the video of whether the mistake is present or not. By adopting a VAD formulation, the model can be trained with weak supervision, requiring only a single label per video. Additionally, to alleviate the problem of data scarcity of incorrect executions, we introduce a multi-robot simulation dataset with controlled temporal errors and real executions for zero-shot sim-to-real evaluation. Our experiments demonstrate that out-of-the-box VLMs lack the explicit temporal reasoning required for this task, whereas our framework successfully detects different types of temporal errors. Project: https://ropertunizar.github.io/TIMID/

Vision-Language Navigation in Continuous Environments (VLN-CE) requires agents to learn complex reasoning from long-horizon human interactions. While Multi-modal Large Language Models (MLLMs) have driven recent progress, current training paradigms struggle to balance generalization capability, error recovery and training stability. Specifically, (i) policies derived from SFT suffer from compounding errors, struggling to recover from out-of-distribution states, and (ii) Reinforcement Fine-Tuning (RFT) methods e.g. GRPO are bottlenecked by sparse outcome rewards. Their binary feedback fails to assign credit to individual steps, leading to gradient signal collapse in failure dominant batches. To address these challenges, we introduce Step-Aware Contrastive Alignment (SACA), a framework designed to extract dense supervision from imperfect trajectories. At its core, the Perception-Grounded Step-Aware auditor evaluates progress step-by-step, disentangling failed trajectories into valid prefixes and exact divergence points. Leveraging these signals, Scenario-Conditioned Group Construction mechanism dynamically routes batches to specialized resampling and optimization strategies. Extensive experiments on VLN-CE benchmarks demonstrate that SACA achieves state-of-the-art performance.

Recent embodied navigation approaches leveraging Vision-Language Models (VLMs) demonstrate strong generalization in versatile Vision-Language Navigation (VLN). However, reliable path planning in complex environments remains challenging due to insufficient spatial awareness. In this work, we introduce SPAN-Nav, an end-to-end foundation model designed to infuse embodied navigation with universal 3D spatial awareness using RGB video streams. SPAN-Nav extracts spatial priors across diverse scenes through an occupancy prediction task on extensive indoor and outdoor environments. To mitigate the computational burden, we introduce a compact representation for spatial priors, finding that a single token is sufficient to encapsulate the coarse-grained cues essential for navigation tasks. Furthermore, inspired by the Chain-of-Thought (CoT) mechanism, SPAN-Nav utilizes this single spatial token to explicitly inject spatial cues into action reasoning through an end-to end framework. Leveraging multi-task co-training, SPAN-Nav captures task-adaptive cues from generalized spatial priors, enabling robust spatial awareness to generalize even to the task lacking explicit spatial supervision. To support comprehensive spatial learning, we present a massive dataset of 4.2 million occupancy annotations that covers both indoor and outdoor scenes across multi-type navigation tasks. SPAN-Nav achieves state-of-the-art performance across three benchmarks spanning diverse scenarios and varied navigation tasks. Finally, real-world experiments validate the robust generalization and practical reliability of our approach across complex physical scenarios.

VLA models encode visual observations as 2D patch tokens with no intrinsic geometric structure. We introduce GST-VLA with two contributions. First, the Gaussian Spatial Tokenizer (GST) converts frozen dense depth and frozen semantic patch features into $N_g{=}128$ anisotropic 3D Gaussian primitives, each parameterized by a metric residual mean $μ\in \mathbb{R}^3$, log-scale covariance $\log σ\in \mathbb{R}^3$, and learned opacity $α\in (0,1)$. The covariance eigenstructure encodes local surface orientation, and opacity provides per-primitive geometric confidence, both inaccessible from scalar depth. Spatial attention pooling with learned queries concentrates the fixed token budget on geometrically salient regions rather than distributing uniformly. Second, 3D Depth-Aware Chain-of-Thought (DA-CoT) reasoning supervises four structured intermediate spatial thoughts, covering 3D object grounding, grasp affordance contact geometry, pairwise metric distances, and coarse SE(3) waypoints, as explicit generation targets in the training loss. A cross-attention sublayer at every VLM transformer block provides direct access to the raw 256-primitive Gaussian field during DA-CoT generation. A 300M-parameter flow-matching action expert with mixture-of-experts feedforward sublayers decodes 7-DoF delta action chunks via conditional ODE integration, conditioned on both VLM hidden states and DA-CoT outputs through dual cross-attention. Trained with composite $\mathcal{L}_\mathrm{flow} + \mathcal{L}_\mathrm{CoT} + \mathcal{L}_\mathrm{depth}$ across three progressive stages, GST-VLA achieves 96.4% on LIBERO (+2.0%), and 80.2% on SimplerEnv (+5.4%). Ablations isolate the contribution of each GST component, each DA-CoT thought, and each training stage, confirming independent and synergistic gains concentrated on precision demanding tasks.

Robots in shared workspaces must interpret human actions from partial, ambiguous observations, where overconfident early predictions can lead to unsafe or disruptive interaction. This challenge is amplified in egocentric views, where viewpoint changes and occlusions increase perceptual noise and ambiguity. As a result, downstream human-robot interaction modules require not only an action hypothesis but also a trustworthy estimate of confidence under partial observation. Recent vision-language model-based approaches have been proposed for short-term action recognition due to their open-vocabulary and context-aware reasoning, but their uncertainty reliability in the temporal-prefix regime is largely uncharacterized. We present the first systematic evaluation of uncertainty in vision-language model-based short-term action recognition for human-robot interaction. We introduce a temporal-prefix evaluation protocol and metrics for calibration and selective prediction. We also characterize miscalibration patterns and failure modes under partial observations. Our study provides the missing reliability evidence needed to use vision-language model predictions in confidence-gated human-robot interaction modules.

Autonomous navigation in highly constrained environments remains challenging for mobile robots. Classical navigation approaches offer safety assurances but require environment-specific parameter tuning; end-to-end learning bypasses parameter tuning but struggles with precise control in constrained spaces. To this end, recent robot learning approaches automate parameter tuning while retaining classical systems' safety, yet still face challenges in generalizing to unseen environments. Recently, Vision-Language-Action (VLA) models have shown promise by leveraging foundation models' scene understanding capabilities, but still struggle with precise control and inference latency in navigation tasks. In this paper, we propose Adaptive Planner Parameter Learning from Vision-Language-Action Model (applv). Unlike traditional VLA models that directly output actions, applv leverages pre-trained vision-language models with a regression head to predict planner parameters that configure classical planners. We develop two training strategies: supervised learning fine-tuning from collected navigation trajectories and reinforcement learning fine-tuning to further optimize navigation performance. We evaluate applv across multiple motion planners on the simulated Benchmark Autonomous Robot Navigation (BARN) dataset and in physical robot experiments. Results demonstrate that applv outperforms existing methods in both navigation performance and generalization to unseen environments.

The rapid advancement of Embodied Intelligence has opened transformative opportunities in healthcare, particularly in physical therapy and rehabilitation. However, critical challenges remain in developing robust embodied healthcare solutions, such as the lack of standardized evaluation benchmarks and the scarcity of open-source multimodal acupoint massage datasets. To address these gaps, we construct MedMassage-12K - a multimodal dataset containing 12,190 images with 174,177 QA pairs, covering diverse lighting conditions and backgrounds. Furthermore, we propose a hierarchical embodied massage framework, which includes a high-level acupoint grounding module and a low-level control module. The high-level acupoint grounding module uses multimodal large language models to understand human language and identify acupoint locations, while the low-level control module provides the planned trajectory. Based on this, we evaluate existing MLLMs and establish a benchmark for embodied massage tasks. Additionally, we fine-tune the Qwen-VL model, demonstrating the framework's effectiveness. Physical experiments further confirm the practical applicability of the framework.Our dataset and code are publicly available at https://github.com/Xiaofeng-Han-Res/HMR-1.

Accurate pre-contact grasp force selection is critical for safe and reliable robotic manipulation. Adaptive controllers regulate force after contact but still require a reasonable initial estimate. Starting a grasp with too little force requires reactive adjustment, while starting a grasp with too high a force risks damaging fragile objects. This trade-off is particularly challenging for compliant grippers, whose contact mechanics are difficult to model analytically. We propose Exp-Force, an experience-conditioned framework that predicts the minimum feasible grasping force from a single RGB image. The method retrieves a small set of relevant prior grasping experiences and conditions a vision-language model on these examples for in-context inference, without analytic contact models or manually designed heuristics. On 129 object instances, ExpForce achieves a best-case MAE of 0.43 N, reducing error by 72% over zero-shot inference. In real-world tests on 30 unseen objects, it improves appropriate force selection rate from 63% to 87%. These results demonstrate that Exp-Force enables reliable and generalizable pre-grasp force selection by leveraging prior interaction experiences. http://expforcesubmission.github.io/Exp-Force-Website/

Recent vision-language-action (VLA) models have significantly advanced robotic manipulation by unifying perception, reasoning, and control. To achieve such integration, recent studies adopt a predictive paradigm that models future visual states or world knowledge to guide action generation. However, these models emphasize forecasting outcomes rather than reasoning about the underlying process of change, which is essential for determining how to act. To address this, we propose $Δ$VLA, a prior-guided framework that models world-knowledge variations relative to an explicit current-world knowledge prior for action generation, rather than regressing absolute future world states. Specifically, 1) to construct the current world knowledge prior, we propose the Prior-Guided WorldKnowledge Extractor (PWKE). It extracts manipulable regions, spatial relations, and semantic cues from the visual input, guided by auxiliary heads and prior pseudo labels, thus reducing redundancy. 2) Building upon this, to represent how world knowledge evolves under actions, we introduce the Latent World Variation Quantization (LWVQ). It learns a discrete latent space via a VQ-VAE objective to encode world knowledge variations, shifting prediction from full modalities to compact latent. 3)Moreover, to mitigate interference during variation modeling, we design the Conditional Variation Attention (CV-Atten), whichpromotes disentangled learning and preserves the independence of knowledge representations. Extensive experiments on both simulated benchmarks and real-world robotic tasks demonstrate $Δ$VLA achieves state-of-the-art performance while improving efficiency. Code and real-world execution videos are available at https://github.com/JiuTian-VL/DeltaVLA.

Existing data generation methods for robot learning suffer from limited exploration, embodiment gaps, low signal-to-noise ratios, and domain shifts, leading to performance degradation during self-iteration and poor generalization to unseen scenes. To address these challenges, we propose Seed2Scale, a self-evolving data engine with parallel worlds expansion. Starting with as few as four seed demonstrations, Seed2Scale first executes a self-evolution stage driven by a heterogeneous synergy of "small-model collection, large-model evaluation, and target-model learning". Specifically, the lightweight Vision-Language-Action (VLA) model, SuperTiny, serves as a dedicated data collector for robust exploration. Concurrently, a pretrained Vision-Language Model (VLM) functions as a verifier to autonomously score and filter trajectories, supporting stable self-evolution in the evaluated tasks without performance collapse. Furthermore, Seed2Scale introduces a parallel worlds stage, projecting self-evolved trajectories into different environments of the same task to generate more diverse data and enhance adaptability to unseen scenes, including real-world environments. Experimental results demonstrate that Seed2Scale exhibits significant scaling potential: as iterations progress, the success rate of the target model shows a consistent upward trend, significantly outperforming the seed baseline. Notably, Seed2Scale achieves a remarkable 75.38% success rate in zero-shot real-world evaluations, where baseline methods fail completely (0%). Project page: https://terminators2025.github.io/Seed2Scale.github.io

While Vision-Language-Action (VLA) models have demonstrated remarkable success in robotic manipulation, their application has largely been confined to low-degree-of-freedom end-effectors performing simple, vision-guided pick-and-place tasks. Extending these models to human-like, bimanual dexterous manipulation-specifically contact-rich in-hand operations-introduces critical challenges in high-fidelity data acquisition, multi-skill learning, and multimodal sensory fusion. In this paper, we propose an integrated framework to address these bottlenecks, built upon two components. First, we introduce IMCopilot (In-hand Manipulation Copilot), a suite of reinforcement learning-trained atomic skills that plays a dual role: it acts as a shared-autonomy assistant to simplify teleoperation data collection, and it serves as a callable low-level execution primitive for the VLA. Second, we present MoDE-VLA (Mixture-of-Dexterous-Experts VLA), an architecture that seamlessly integrates heterogeneous force and tactile modalities into a pretrained VLA backbone. By utilizing a residual injection mechanism, MoDE-VLA enables contact-aware refinement without degrading the model's pretrained knowledge. We validate our approach on four tasks of escalating complexity, demonstrating doubled success rate improvement over the baseline in dexterous contact-rich tasks.

Recent advances in Visual-Language-Action (VLA) models have shown promising potential for robotic manipulation tasks. However, real-world robotic tasks often involve long-horizon, multi-step problem-solving and require generalization for continual skill acquisition, extending beyond single actions or skills. These challenges present significant barriers for existing VLA models, which use monolithic action decoders trained on aggregated data, resulting in poor scalability. To address these challenges, we propose AtomicVLA, a unified planning-and-execution framework that jointly generates task-level plans, atomic skill abstractions, and fine-grained actions. AtomicVLA constructs a scalable atomic skill library through a Skill-Guided Mixture-of-Experts (SG-MoE), where each expert specializes in mastering generic yet precise atomic skills. Furthermore, we introduce a flexible routing encoder that automatically assigns dedicated atomic experts to new skills, enabling continual learning. We validate our approach through extensive experiments. In simulation, AtomicVLA outperforms $π_{0}$ by 2.4% on LIBERO, 10% on LIBERO-LONG, and outperforms $π_{0}$ and $π_{0.5}$ by 0.22 and 0.25 in average task length on CALVIN. Additionally, our AtomicVLA consistently surpasses baselines by 18.3% and 21% in real-world long-horizon tasks and continual learning. These results highlight the effectiveness of atomic skill abstraction and dynamic expert composition for long-horizon and lifelong robotic tasks. The project page is here.

We present DreamToNav, a novel autonomous robot framework that uses generative video models to enable intuitive, human-in-the-loop control. Instead of relying on rigid waypoint navigation, users provide natural language prompts (e.g. ``Follow the person carefully''), which the system translates into executable motion. Our pipeline first employs Qwen 2.5-VL-7B-Instruct to refine vague user instructions into precise visual descriptions. These descriptions condition NVIDIA Cosmos 2.5, a state-of-the-art video foundation model, to synthesize a physically consistent video sequence of the robot performing the task. From this synthetic video, we extract a valid kinematic path using visual pose estimation, robot detection and trajectory recovery. By treating video generation as a planning engine, DreamToNav allows robots to visually "dream" complex behaviors before executing them, providing a unified framework for obstacle avoidance and goal-directed navigation without task-specific engineering. We evaluate the approach on both a wheeled mobile robot and a quadruped robot in indoor navigation tasks. DreamToNav achieves a success rate of 76.7%, with final goal errors typically within 0.05-0.10 m and trajectory tracking errors below 0.15 m. These results demonstrate that trajectories extracted from generative video predictions can be reliably executed on physical robots across different locomotion platforms.

Large language models have been widely used for robotic task planning, often taking advantage of representations such as Behavior Trees (BTs). Vision-Language Models (VLMs) have extended these works by grounding the generated plans in the observed scene. However, existing methods are either text-only or rely on large proprietary VLMs, while no dataset pairs visual observations and task instructions with executable and ROS2-compatible BTs. We address this gap with a multi-stage teacher pipeline that converts 1,622 Open X-Embodiment episodes into an augmented multimodal BT dataset containing 2,433 examples. We use this dataset to fine-tune compact and open-source VLMs, ranging from 500M to 4B parameters, using parameter-efficient fine-tuning (PEFT). We then evaluate the generated BTs offline in terms of syntactic correctness and by executing them on 15 household tasks in BEHAVIOR-1K. Our best model, Gemma-3 4B, achieves perfect BT validity and an 87% success rate, outperforming Claude Opus 4.8 and approaching GPT-5, while running locally. Finally, our ablation studies show that adding visual observations increases task success from 40% to 87%, while data augmentation increases BT validity from 65% to 100%.

Effective communication is vital in healthcare, especially across language barriers, where non-verbal cues and gestures are critical. This paper presents a privacy-preserving vision-language framework for medical interpreter robots that detects specific speech acts (consent and instruction) and generates corresponding robotic gestures. Built on locally deployed open-source models, the system utilizes a Large Language Model (LLM) with few-shot prompting for intent detection. We also introduce a novel dataset of clinical conversations annotated for speech acts and paired with gesture clips. Our identification module achieved 0.90 accuracy, 0.93 weighted precision, and a 0.91 weighted F1-Score. Our approach significantly improves computational efficiency and, in user studies, outperforms the speech-gesture generation baseline in human-likeness while maintaining comparable appropriateness.

Vision-Language-Action models (VLAs) have shown remarkable progress towards embodied intelligence. While their architecture partially resembles that of Large Language Models (LLMs), VLAs exhibit higher complexity due to their multi-modal inputs/outputs and often hybrid nature of transformer and diffusion heads. This is part of the reason why insights from mechanistic interpretability in LLMs, which explain how the internal model representations relate to their output behavior, do not trivially transfer to VLA counterparts. In this work, we investigate whether VLA internal representations support lightweight behavioral steering without retraining. Across four frontier VLA models, linear observers recover state- and action-relevant information in both autoregressive and transformer--flow-matching architectures, and provide robust directions to causally alter VLA outputs. Building on this, we introduce a controller that minimally modifies representations to place observer predictions within prescribed target intervals. Closed-loop experiments with $π_{0.5}$ in the LIBERO simulator and on DROID hardware demonstrate improved constraint satisfaction while retaining task performance, with only approximately $1\%$ inference overhead. Together, these experiments show that lightweight linear interventions can reliably steer VLA behavior while preserving closed-loop capabilities, enabling alignment with user preferences and task requirements without fine-tuning.

Open-world navigation requires robots to make decisions in complex everyday environments while adapting to flexible task requirements. Conventional navigation approaches often rely on dense 3D reconstruction and hand-crafted goal metrics, which limits their generalization across tasks and environments. Recent advances in vision-language navigation (VLN) and vision-language-action (VLA) models enable end-to-end policies conditioned on natural language, but typically require interactive training, large-scale data collection, or task-specific fine-tuning with a mobile agent. We formulate navigation as a sparse subgoal identification and reaching problem and observe that providing visual anchoring targets for high-level semantic priors enables highly efficient goal-conditioned navigation. Based on this insight, we select visual frontiers as semantic anchors and propose OpenFrontier, a navigation framework that requires no task-specific training or fine-tuning and seamlessly integrates diverse vision-language prior models. OpenFrontier enables efficient navigation with a lightweight system design, without dense 3D semantic mapping, task-specific policy training, or model fine-tuning. We evaluate OpenFrontier across multiple navigation benchmarks and demonstrate strong zero-shot performance, as well as effective real-world deployment on a mobile robot.

The increasing demand for sustainable textile recycling requires robust automation solutions capable of handling deformable garments and detecting foreign objects in cluttered environments. This work presents a digital twin driven robotic sorting system that integrates grasp prediction, multi modal perception, and semantic reasoning for real world textile classification. A dual arm robotic cell equipped with RGBD sensing, capacitive tactile feedback, and collision-aware motion planning autonomously separates garments from an unsorted basket, transfers them to an inspection zone, and classifies them using state of the art Visual Language Models (VLMs). We benchmark nine VLM s from five model families on a dataset of 223 inspection scenarios comprising shirts, socks, trousers, underwear, foreign objects (including garments outside of the aforementioned classes), and empty scenes. The evaluation assesses per class accuracy, hallucination behavior, and computational performance under practical hardware constraints. Results show that the Qwen model family achieves the highest overall accuracy (up to 87.9 %), with strong foreign object detection performance, while lighter models such as Gemma3 offer competitive speed accuracy trade offs for edge deployment. A digital twin combined with MoveIt enables collision aware path planning and integrates segmented 3D point clouds of inspected garments into the virtual environment for improved manipulation reliability. The presented system demonstrates the feasibility of combining semantic VLM reasoning with conventional grasp detection and digital twin technology for scalable, autonomous textile sorting in realistic industrial settings.